Thank you for providing your code. After reading your code, I have several questions and I really appreciate it if you could answer these questions for me.

(1) In the paper, it is said that a threshold is used to select components of W_p (set to 1.25 in experiments), but I did not find any trace of it in the codes.
(2) In the code below, I don't understand the reason for setting the ratio for class p to 10 when the denominator is zero, why it should be 10 and not some other numbers? Is there any other empirical observation to support that?

FEMNIST-monitor/models/Fed.py in whole_determination()
if aux_other_sum != 0:
            res = abs(aux_sum) / abs(aux_other_sum)  
        else:  
            res = 10  

(3) In the code, Np, the predicted number of classes p, is computed by the code below. It is confusing that the code computes the average value of Ra_pi by summing up all the nominators and denominators respectively and divides those two summations to get the ratio Ra_p for class p. Could you please explain why?

FEMNIST-monitor/models/Fed.py in whole_determination()
for it in range(26):
        cc_class = it
        aux_sum = 0
        aux_other_sum = 0
        layer = 1
        for i in range(pos.shape[1]):
            for j in range(pos.shape[2]):
                if pos[cc_class, i, j] == 1:
                    temp = []
                    last = w_glob_last['resnet.fc{}.weight'.format(layer)].cpu().numpy()[i, j]
                    cc = cc_net[cc_class]['resnet.fc{}.weight'.format(layer)].cpu().numpy()[i, j]
                    for p in range(len(cc_net)):
                        temp.append(cc_net[p]['resnet.fc{}.weight'.format(layer)].cpu().numpy()[i, j] - last)
                    temp = np.array(temp)
                    temp = np.delete(temp, cc_class)
                    temp_ave = np.sum(temp) / (len(cc_net) - 1)
                    aux_sum += cc - last
                    aux_other_sum += temp_ave
        if aux_other_sum != 0:
            res = abs(aux_sum) / abs(aux_other_sum)
        else:
            res = 10

(4) In the following code, why the predicted number of class p should be less than 1.5 times the average number of samples for each class?

FEMNIST-monitor/models/Fed.py in monitoring()

if 0 < res_temp < num_samples * 1.5 / num_class:
    temp_res.append(res_temp)

Thank you for your attention and any help/answers will be highly appreciated.

Hi, I see your work is using image dataset. How can I use it on a tabular format csv dataset such as: https://www.kaggle.com/wenruliu/adult-income-dataset
Thanks!

Hello，I have read your paper, Addressing Class Imbalance in Federated Learning. It is really a wonderful idea! Will you open your source code? I'm looking forward to learning it! Thanks a lot!